Fog detection using an artificial neural network
(2023) In Master's Theses in Mathematical Sciences FMSM01 20231Mathematical Statistics
- Abstract
- This project studies a method of image-based fog detection directly from a camera without using the transmissometer.
Fog can be detected using transmissometers which could be a very costly approach. This thesis presents an image-based approach for fog detection using Artificial Neural networks. A neural network model will be used to classify images either into two classes, whether it contains fog or not, or into multiple fog intensity classes, based on visibility derived from Koschmieder's Law. By applying neural network model foggy weather can be detected directly from surveillance cameras and makes fairly accurate predictions, which is useful in many aspects of industry and life.
The goal of the thesis is to improve on... (More) - This project studies a method of image-based fog detection directly from a camera without using the transmissometer.
Fog can be detected using transmissometers which could be a very costly approach. This thesis presents an image-based approach for fog detection using Artificial Neural networks. A neural network model will be used to classify images either into two classes, whether it contains fog or not, or into multiple fog intensity classes, based on visibility derived from Koschmieder's Law. By applying neural network model foggy weather can be detected directly from surveillance cameras and makes fairly accurate predictions, which is useful in many aspects of industry and life.
The goal of the thesis is to improve on previous methods that have used neural networks for similar tasks. To achieve the goal, the work starts with creating a suitable dataset for training and validation based on a combination of real images and images generated from Unreal Engine or other simulation scrips. And then, create an effective model based on neural network with the dataset to solve the classification problem. Finally, validate the completed model and implement it on a real surveillance camera to test the actual performance, the model will be retrained if necessary according to the test result. (Less) - Popular Abstract
- Popular Science Summary
Fog detection using an artificial neural network
Tiancheng Ma & Quanwei Li
In surveillance camera industry, the presence of fog causes various problems in scenarios such as sector security, road safety and self-driving. By removing the fog there are methods such as using and defog filter and it can be controlled manually in the VMS. To automate the fog detection and the control the defog filter, there are some developed methods using computer vision and the usage of Artificial Intelligence (AI), which could be used to let the machine know when to turn on or turn off the defog filter to ensure a better visibility while keeping up the image quality. In this thesis we will employ a form of AI known as “Artificial... (More) - Popular Science Summary
Fog detection using an artificial neural network
Tiancheng Ma & Quanwei Li
In surveillance camera industry, the presence of fog causes various problems in scenarios such as sector security, road safety and self-driving. By removing the fog there are methods such as using and defog filter and it can be controlled manually in the VMS. To automate the fog detection and the control the defog filter, there are some developed methods using computer vision and the usage of Artificial Intelligence (AI), which could be used to let the machine know when to turn on or turn off the defog filter to ensure a better visibility while keeping up the image quality. In this thesis we will employ a form of AI known as “Artificial Neural Network” (ANN) that detects fog from the video stream and configures the defog filter automatically.
Our method includes two main parts, an image pre-processing pipeline, and an ANN predictor. The first part is to clear the unusable information from the original images by cutting and then recreate the images of different sizes to the same size which are suitable to be used by the predictor. The second part is to make a predictor powered by an ANN. It works by taking an image as an input, and it will output a prediction in the form of the probability indicating how “likely” the input belongs to each category. Finally, the ANN will predict the input’s class by choosing the most possible category based on the probability values.
An ANN is a digital network structure that is made by different layers which connects to each other by a decided order, and each layer will transfer its output to layers of its next level. Each layer has the same/different structures and specifications. These layers will optimize themselves under each step of the training process to find the best solution for the classification problem.
The ANN predictor was implemented and tested using an image set with more various sceneries to validate its actual performance. Once the model is proven successful, it will be implemented a video stream to do real-time fog detection and finally implement it on a pan-till-zoom camera for live video stream testing. This whole training process can be repeated several times to become better and better. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9126782
- author
- Li, Quanwei LU and Ma, Tiancheng
- supervisor
- organization
- course
- FMSM01 20231
- year
- 2023
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Machine Learning, Deep Learning, Image Analysis, Computer Vision
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUNFMS-3120-2023
- ISSN
- 1404-6342
- other publication id
- 2023:E48
- language
- English
- id
- 9126782
- date added to LUP
- 2023-06-19 13:04:38
- date last changed
- 2023-06-19 15:43:49
@misc{9126782,
abstract = {{This project studies a method of image-based fog detection directly from a camera without using the transmissometer.
Fog can be detected using transmissometers which could be a very costly approach. This thesis presents an image-based approach for fog detection using Artificial Neural networks. A neural network model will be used to classify images either into two classes, whether it contains fog or not, or into multiple fog intensity classes, based on visibility derived from Koschmieder's Law. By applying neural network model foggy weather can be detected directly from surveillance cameras and makes fairly accurate predictions, which is useful in many aspects of industry and life.
The goal of the thesis is to improve on previous methods that have used neural networks for similar tasks. To achieve the goal, the work starts with creating a suitable dataset for training and validation based on a combination of real images and images generated from Unreal Engine or other simulation scrips. And then, create an effective model based on neural network with the dataset to solve the classification problem. Finally, validate the completed model and implement it on a real surveillance camera to test the actual performance, the model will be retrained if necessary according to the test result.}},
author = {{Li, Quanwei and Ma, Tiancheng}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{Fog detection using an artificial neural network}},
year = {{2023}},
}